Overview / Usage

CAE simulations remain computationally expensive, necessarily so to ensure highest accuracy as product development moves towards exclusive virtual prototyping. CAE algorithms continue to improve their inherent multi-core and multi-node parallelism, but even reduced approximated simulations can take minutes to hours to run. Testing a range of scenarios, even for small changes in the input model, is constrained by available resources. And although optimisation tools help in generating and executing models within a defined design space there are few options that allow the engineer to practically explore model changes interactively.

Fujitsu’s “AI-Solver” is a data-driven technique that can learn from physics-based simulation to instantly predict the principal field distribution of within a 3D space. It uses a deep learning framework to learn the response of a system from simulation data generated on arbitrarily-shaped geometries.

The project is a join work between Fujitsu Systems Europe and Fujitsu Labs (Europe). Team internal members are Serban Georgescu (FLE), Ahmed Al-Jarro (FLE), Ian Godfrey (FSE), Michael Dussere (FSE), Pierre Lagier (FSE).

Methodology / Approach

The principle of the AI-Solver approach is that a neural network can be trained to produce results that are sufficiently close to those from direct numerical simulation. And that with enough training these results can be accurate enough for data-driven methods to become viable in the design engineering ecosystem. (Clearly there is strong corollary with the way in which human engineers acquire experience and are then able to estimate quite effectively the likely outcome of a given new design.) Another requirement must also be that the neural network algorithm performs substantially faster than numerical simulation to be of enough value to compensate for any imprecision, or rather the absence of formal justification behind a given result.

​

Technologies Used

Setup of the AI-Solver deep neural network (DNN) comprises the essential well-known stages of training, testing and inference. Intel hardware technologies are leveraged in all phases of executing the DNN, with large-scale parallelism during the limited-duration training/testing, and now particularly the new vector neural network instructions (VNNI) for dynamic inference of the resulting AI model.

As such we heavily rely on Intel technologies, both hardware and software, the fully integrated Intel python environment including MKL as well as TensorFlow.

​

Other links

• Wie Fujitsu KI-Simulation in Echtzeit ermöglicht